Measure and understand
Commerce and AI Shelf
Measure product and category appearances with target-specific buyer prompts.
Commerce’s catalog is the starting point: choose a product or category, then inspect its AI Shelf, competitors and buyer prompts. Site Health crawls fill the catalog for businesses that sell a product catalog; any business can import a CSV.
Review the catalog
Check the selected target’s identity and website page. Review competitor candidates before approving them. A discovered candidate does not enter measurement until it is approved.
Product and category targets represent different buying contexts, so keep the target explicit throughout your analysis.
Prepare buyer prompts
Create or review prompts for the selected target. Generated and manual Commerce prompts remain disabled until explicit approval.
When launching an audit, select the approved prompts and review engines, repetitions and estimated cost. The resulting audit freezes the relevant catalog and competitor identities; later catalog edits do not reinterpret old answers.
Read AI Shelf metrics
| Metric | Meaning |
|---|---|
| Product Visibility | Successful target executions with an owned appearance, divided by successful target executions |
| Share of Shelf | Owned recognized slots, divided by all recognized slots |
| Average Shelf Position | Mean owned rank where recommendations have an explicit order |
| First-Position Win Rate | Eligible ordered executions whose first slot is owned, divided by eligible ordered executions |
A recommendation in prose or an unordered bullet list does not supply a supported rank. Average position can therefore be unavailable even when an owned product appears. An answer that names the same product several times recommends it once. When no answer for the target succeeded, visibility is unavailable rather than zero.
Each metric shows when it was measured and how many answers and recommendations it rests on.
Inspect the recommendation
Who held the shelf lists what the latest audit’s answers recommended for the target: your products, approved competitors and products seen only in AI answers, with how many answers named each and their best position. Product identity, merchant domain and citation URL are separate observations.
If a result is unresolved, retain that uncertainty. It is better to report a limited observed sample than to invent a complete product ranking.
Decide what to improve
The target shows one next step, and its open Actions once it is measured. Open an Action to prepare a reviewable change, alone or with the Agent. Publishing and later measurement remain separate decisions.